v0.2.2

Fix diarization OOM on long audio files

Breaking change: None — drop-in fix.

What changed

• diarize() now uses model.generate_stream() (chunked streaming) instead of model.generate() (full-file, O(n²) memory)
• Added _merge_diarization_segments() for cross-chunk speaker segment merging
• 9 new unit tests + 2 integration tests with 82-min public domain podcast

Before/After

• Before: 48-min file → [metal::malloc] Attempting to allocate 83404284800 bytes (OOM)
• After: 48-min file → 12.6s diarization (289 chunks, 712 segments, 4 speakers)

Other fixes

• Fixed test_cli.py hardcoded version string
• Updated CLAUDE.md with current line counts